OWASP Top 10 for LLM and Agentic Applications

OWASP's two community risk lists for generative AI: the LLM Top 10 (2023-24, 2025 and 2026 editions) and the Agentic Top 10 (ASI01-ASI10), what each entry means, how the editions differ, and how to use them in a design or code review.

Contents

The OWASP Top 10 for LLM Applications is a ranked, community-maintained list of the ten most critical security risks in applications built on large language models, published by the OWASP GenAI Security Project. It has had three editions: 2023-24 (v1.1, dated 2024-04-11), 2025 (published 2024-11-17) and 2026 (v1.0, published 2026-08-03) (2023-24, 2025 pages, 2026). Its companion, the OWASP Top 10 for Agentic Applications (labelled 2026, released 2025-12-09, entries ASI01 to ASI10), covers systems where the model plans, calls tools, keeps memory and talks to other agents (Agentic Top 10). Both are awareness documents: they name and rank risk classes, but they are not testable requirements (Notebook OWASP). This page covers the lists as frameworks. For the threat model and controls risk by risk, see LLM security risks.

How the lists fit together

The 2026 LLM list treats the model as a component of an application. Once the model becomes an actor, with tools it can call, memory across sessions and downstream consequences, the risk is routed to the Agentic Top 10 (Help Net Security). The 2026 document maps its entries to NIST, MITRE ATLAS, CWE and the Agentic Top 10 (2026). Alongside the 2026 release OWASP presented the Agent Control Standard (ACS) v0.1, a separate specification for enforcing controls on agents at runtime (CSA).

Synthesis: lineage and scope of the OWASP GenAI documents.

flowchart LR
  A["LLM Top 10 2023-24<br/>v1.1, 2024-04-11"] --> B["LLM Top 10 2025<br/>2024-11-17"]
  B --> C["LLM Top 10 2026<br/>v1.0, 2026-08-03"]
  D["Agentic Top 10 (ASI) 2026<br/>2025-12-09"] -->|"cross-mapped"| C
  C -->|"model as component"| R["Design / code review"]
  D -->|"model as actor"| R
  E["Agent Control Standard v0.1<br/>runtime enforcement"] --> R
  F["ASVS / WSTG<br/>testable requirements"] --> R

The 2026 LLM list, entry by entry

IDRiskOne-line meaning
LLM01:2026Prompt InjectionUntrusted text (user, documents, tool output, memory, images or audio) changes model behaviour; see prompt injection
LLM02:2026Sensitive Information DisclosureThe app reveals PII, secrets or data beyond the user’s rights, including through retrieval and logs; see PII and data privacy
LLM03:2026Excessive AgencyToo much functionality, permission or autonomy, so a manipulated output does real damage
LLM04:2026Supply ChainTampered or substituted weights, adapters, templates, unsafe formats, hallucinated packages
LLM05:2026Data and Model PoisoningCorrupted pre-training, fine-tuning, adapter or embedding data plants bias or backdoors
LLM06:2026Unbounded ConsumptionCost and availability abuse, amplified by reasoning models and agent loops
LLM07:2026MisinformationWrong output relied on, now often by automated downstream steps; see hallucination
LLM08:2026Hidden Context ExposureLeakage of system prompts, tool schemas, policy text, retrieved documents, memory and app state
LLM09:2026Vector and Embedding WeaknessesCross-tenant leakage, retrieval manipulation and embedding inversion in vector databases
LLM10:2026Improper Output HandlingModel output reaches HTML, SQL, shells, terminals (ANSI escapes) or auto-fetching renderers unvalidated

Sources: order and scope changes from 2026, Help Net Security and CSA. The official PDF was not fetched directly; three secondary summaries agree on the order.

Edition differences

Risk theme2023-24 (v1.1)20252026
Prompt injectionLLM01LLM01LLM01
Sensitive information disclosureLLM06LLM02LLM02
Excessive agencyLLM08LLM06LLM03
Supply chainLLM05 Supply Chain VulnerabilitiesLLM03LLM04
PoisoningLLM03 Training Data PoisoningLLM04 Data and Model PoisoningLLM05
Resource abuseLLM04 Model Denial of Service; LLM10 Model TheftLLM10 Unbounded ConsumptionLLM06
Wrong output relied onLLM09 OverrelianceLLM09 MisinformationLLM07
Hidden instructions and context(none)LLM07 System Prompt LeakageLLM08 Hidden Context Exposure
RAG and vector stores(none)LLM08 Vector and Embedding WeaknessesLLM09
Output handlingLLM02 Insecure Output HandlingLLM05 Improper Output HandlingLLM10
PluginsLLM07 Insecure Plugin Design(no standalone entry)(none)

Sources: 2023-24, 2025, 2026.

2023-24 to 2025. The 2025 edition added System Prompt Leakage (incidents showed prompts leak) and Vector and Embedding Weaknesses (RAG had become central). It widened Training Data Poisoning into Data and Model Poisoning, recast Overreliance as Misinformation, and replaced denial of service with Unbounded Consumption, which also covers runaway cost. Excessive Agency grew in scope as agent architectures spread (Qualys). LLM10:2025 lists model extraction via API and functional model replication among its examples, and LLM09:2025 treats overreliance as the amplifier of misinformation (2025 pages). Synthesis: Model Theft therefore lives on inside Unbounded Consumption. Insecure Plugin Design’s concerns are spread across Excessive Agency (minimise extensions and their permissions) and Supply Chain (third-party components) (2025); OWASP’s risk pages do not state these mappings themselves.

2025 to 2026. Eight of ten positions changed. The movements that matter for reviewers:

  • Method. Ranking is 75% practitioner vote and 25% evidence from 6,639 real incidents (Help Net Security, CSA).
  • Excessive Agency, #6 to #3. Both votes and incidents point to agent deployments as where damage happens (Help Net Security).
  • Unbounded Consumption, #10 to #6, reframed as cost asymmetry from reasoning models and loops; Misinformation, #9 to #7, because incident data ranked it far above the vote (2026, Help Net Security).
  • System Prompt Leakage renamed Hidden Context Exposure, widening confidentiality from the prompt to everything placed in context (CSA).
  • Improper Output Handling, #5 to #10, while gaining terminal and auto-fetching renderer sinks (2026).
  • Prompt Injection stays #1 despite few public incidents, which OWASP attributes to a “defense effect”: heavy mitigation hides the underlying risk (Help Net Security).

The Agentic Top 10 (ASI01-ASI10)

Built by OWASP’s Agentic Security Initiative with more than 100 contributors, around two principles: least agency (only the autonomy, tools and permissions the task needs) and strong observability (you can see what agents do, why, and with which tools) (NeuralTrust, Agentic Top 10).

IDRiskWhat goes wrongTypical control
ASI01Agent Goal HijackInjected content redirects the agent’s goal or planTreat all natural-language input as untrusted; approve goal changes
ASI02Tool Misuse and ExploitationAn authorised tool is used unsafelyPer-tool least privilege, policy check between planner and tool, budgets
ASI03Identity and Privilege AbuseInherited or delegated credentials reach too farShort-lived, task-scoped tokens per agent; see agent auth
ASI04Agentic Supply Chain VulnerabilitiesA loaded tool, prompt, plugin or MCP server is maliciousSigning, AI-BOM, pinning and allowlists, kill switch
ASI05Unexpected Code ExecutionText becomes executed codeSandboxes, no eval, validation gate
ASI06Memory and Context PoisoningCorrupted memory, summaries or retrieval stores bias later stepsScan writes, segment per user, provenance, rollback
ASI07Insecure Inter-Agent CommunicationSpoofed, replayed or tampered agent messages (A2A)Mutual auth, signed messages, nonces, typed schemas
ASI08Cascading FailuresOne fault spreads across agents and workflowsIndependent policy engine, circuit breakers, rate limits
ASI09Human-Agent Trust ExploitationA fluent agent steers human approvalsPlain-language risk summaries, log shown vs executed; see human-in-the-loop
ASI10Rogue AgentsAn agent drifts while each action looks legitimateBehavioural baselines, owner and expiry per agent, tested kill switch

Source: NeuralTrust, names cross-checked with Agentic Top 10; the official PDF was not fetched.

Each agentic risk maps to one to four entries of the 2025 LLM list (NeuralTrust), and an OWASP governance report found prompt injection maps to six of the ten agentic categories (CSA). Synthesis: ASI01 and ASI06 are prompt injection (LLM01) seen from the agent’s side, one acting on the current goal and one stored for later. ASI02, ASI03 and ASI05 split Excessive Agency (LLM03:2026) into tool, identity and execution. ASI07, ASI08 and ASI10 have no real LLM-list equivalent, because they only exist once there are several agents or long-running autonomy.

Comparisons

  • LLM list vs Agentic list. Use the LLM list for any feature that calls a model. Add the Agentic list as soon as the model selects tools, writes memory or hands off to another agent (Help Net Security).
  • Top 10 vs ASVS/WSTG. Top 10 lists build awareness; ASVS gives verifiable requirements and WSTG gives test procedures (Notebook OWASP).
  • Top 10 vs Agent Control Standard. The Top 10 says what can go wrong; ACS v0.1 specifies runtime enforcement: guardian-agent enforcement points, OpenTelemetry and OCSF tracing, and an Agent Bill of Materials. Its roadmap reaches deny/modify enforcement across A2A and MCP only at v3, so pilot it rather than deploy it (CSA).
  • Top 10 vs design rules. The lethal trifecta and Meta’s Rule of Two are design constraints you can check per feature, not risk categories (Willison, Meta).

Limits and failure modes

  • IDs are not stable across editions. LLM06 means Excessive Agency in 2025 and Unbounded Consumption in 2026; LLM07 means System Prompt Leakage, then Misinformation. Always write the year suffix (LLM03:2026) in tickets, findings and tool configs (2025, 2026).
  • The ranking is mostly opinion. Three quarters of the weight is a practitioner vote, and the incident data is new (Help Net Security). CSA advises building programmes on the risk categories, not one year’s order (CSA).
  • Global rank is not your risk. Synthesis: a read-only internal chatbot and a payment agent face the same list with very different orderings. Rank by your data, tools and outbound channels.
  • Categories overlap. One exfiltration chain can be LLM01, LLM02, LLM03 and LLM10 at once. Synthesis: record the root cause once and tag the other IDs, or the same issue gets counted four times.
  • Coverage is not prevention. No foolproof defence against prompt injection is known (2025); the leads’ advice is to assume the model will be fooled and design so nothing important breaks (Help Net Security).
  • Secondary-source caveat (as of 2026-11). The 2026 order and the ASI definitions on this page come from consistent secondary summaries, not the official PDFs.

Using the lists in a review

Synthesis: a review procedure built from the controls in the sources above.

  1. Scope. Draw the data flow: inputs (user, documents, web, tool results, memory), the model, outputs and their sinks, tools and their credentials. Decide whether the feature is a component (LLM list only) or an actor (add ASI).
  2. Trifecta check first. If a single session combines private data, untrusted content and an outbound channel, flag it before going item by item (Willison, Meta).
  3. Walk each entry as a question with evidence. For example: LLM03/ASI02-03: which tools, which scopes, whose identity, which actions need approval? LLM08: is any secret, limit or authorisation rule in context that the code assumes the model will keep? LLM10/ASI05: where does output go, and is it escaped, schema-validated (structured outputs) or sandboxed? LLM06: are turns, tokens and spend capped per user? ASI06: who can write to memory and the vector store, and can a write be rolled back?
  4. Prefer deterministic controls. Authorisation, secrets and limits belong in code, hooks and middleware and downstream systems, not in prompts (2025). Classifier-style guardrails are a second layer.
  5. Test. Turn findings into red-team cases (LLM red teaming, OWASP red teaming guide). promptfoo ships framework presets owasp:llm (per item, e.g. owasp:llm:01) and owasp:agentic (promptfoo docs, plugins). Check which edition a preset’s numbering follows before mapping results.
  6. Record and re-baseline. Log findings against year-suffixed IDs, give each item an owner (Notebook OWASP), and re-rank existing risk registers when a new edition lands (CSA). Feed the result into AI governance and compliance evidence.

Sources

Suggest changes →